Trustworthy AI in administration: small models, big impact

Small language models, or SLMs for short, are comparatively compact language models with fewer than ten billion parameters. In simple terms, parameters can be described as manipulated variables with which a model learns language. The fewer parameters there are, the more specialized and limited the model’s capabilities.

Large Language Models, or LLMs for short, on the other hand, have significantly more parameters. There are often seventy billion or more, sometimes even several trillion. These models are designed to cover many different language patterns, topics and contexts simultaneously.

There are also differences in origin. Smaller generative models are often created in open research approaches by large technology groups such as Meta or as part of European research and cooperation projects. Large models are predominantly developed by commercial providers, operated centrally and offered as a service.

The difference in size is not a judgment of quality. It describes different strategic tools, each with specific strengths and weaknesses.

Trustworthy AI starts with the right choice of model.

Small language models are tailored to clearly defined tasks. They are particularly suitable for text summaries, categorizations, simple translations or the structured reproduction of content from existing documents.

A key advantage lies in the form of operation. Many SLMs can be operated within your own IT environment, in some cases even completely offline. This makes it much easier to comply with data protection, IT security and regulatory requirements and increases control over data flows and processing.

Another point is technical reliability. Small models are less prone to so-called hallucinations. As they have little knowledge of the world of their own, they rely more heavily on the content provided. If they are specifically combined with files, guidelines or judgments, they often deliver more objective, comprehensible and verifiable results in practice.

Specialization also comes with limitations. SLMs only have limited knowledge of their own, which may also be incorrect or outdated. Without well-structured and checked documents, there is hardly any added value.

Small models also require more organizational preparation. Selection, adaptation, integration and accompanying processes must be consciously designed. SLMs are not universal systems, but precise tools for clearly defined use cases.

Large language models impress with their versatility. They can take on a variety of different tasks without special adaptation, process large amounts of text and formulate linguistically fluent, explanatory answers.

Thanks to their extensive pre-trained general knowledge, they can be used quickly, are usually available immediately and often answer informal questions reliably. For organizations looking for a standardized tool for many different deployment scenarios, this may initially seem attractive.

However, large models entail structural risks. Running costs are often high as usage volumes or text-based billing are used. This makes long-term budget planning difficult.

There are also data protection and control issues. Many LLMs are operated exclusively externally, which means that data sovereignty is only possible to a limited extent. At the same time, the high computing requirements lead to considerable energy and CO₂ consumption.

In terms of content, there is another critical aspect. The extensive acquired knowledge of large models means that they can produce plausible-sounding but technically incorrect content. This tendency to hallucinate is particularly problematic in legally sensitive contexts such as administration and justice.

For public administrations and the judiciary, a clear principle applies: it is not the model that is decisive, but the specific problem. Those who want to solve clearly defined tasks, such as structured work with existing documents, are often better advised to use small generative models.

SLMs are easier to operate in compliance with data protection regulations and are most effective when they are specifically based on verified content. This makes their use sustainable, controllable and trustworthy. Although the initial outlay is higher, the result is a long-term, stable and reliable solution.

Large models are suitable when a deliberate multifunctional tool is sought and dependencies are accepted. In both cases, digital sovereignty is crucial. The central question is always: who controls data, costs and further development?

Westernacher Solutions supports public organizations in precisely these considerations. The approach is not technology-driven, but responsibility-oriented. The Competence Center AI combines technical possibilities with legal, organizational and ethical requirements.

The focus is on sustainable, comprehensible AI applications instead of short-term hypes. This creates a resilient bridge between technological innovation, regulation and the public mandate.

In practice, Westernacher Solutions is already making targeted use of small generative language models, for example in chatbot solutions for churches or for extracting information from documents. The focus here is not on free conversation, but on structured support in everyday working life. This includes finding relevant information, summarizing documents in a comprehensible way and finding your way through complex texts.

The platform www.trusted-llm.de was also set up to provide decision-makers with guidance at this early stage. Various language models can be compared there along key public sector issues.

• How confidently can a model be operated?
• How transparent are the origin and framework conditions?
• And how well does it fit in with legal, organizational and long-term requirements?

Trusted-llm.de makes it clear that there are already real choices today, even beyond a few large providers. The comparison creates transparency without predetermination and supports administrations in making well-founded and responsible AI decisions.